Time Series & Recurrent Neural Networks

ACTL3143 & ACTL5111 Deep Learning for Actuaries

Author

Patrick Laub

Show the package imports 
import random
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error

from keras.models import Sequential
from keras.layers import Dense, Input
from keras.callbacks import EarlyStopping

Time Series

Tabular data vs time series data

Tabular data

We have a dataset \{ \boldsymbol{x}_i, y_i \}_{i=1}^n which we assume are i.i.d. observations.

Brand Mileage # Claims
BMW 101 km 1
Audi 432 km 0
Volvo 3 km 5
\vdots \vdots \vdots

The goal is to predict the y for some covariates \boldsymbol{x}.

Time series data

Have a sequence \{ \boldsymbol{x}_t, y_t \}_{t=1}^T of observations taken at regular time intervals.

Date Humidity Temp.
Jan 1 60% 20 °C
Jan 2 65% 22 °C
Jan 3 70% 21 °C
\vdots \vdots \vdots

The task is to forecast future values based on the past.

Attributes of time series data

  • Temporal ordering: The order of the observations matters.
  • Trend: The general direction of the data.
  • Noise: Random fluctuations in the data.
  • Seasonality: Patterns that repeat at regular intervals.
Note

Question: What will be the temperature in Berlin tomorrow? What information would you use to make a prediction?

Australian financial stocks

stocks = pd.read_csv("aus_fin_stocks.csv")
stocks
Date ANZ ASX200 BOQ CBA NAB QBE SUN WBC
0 1981-01-02 1.588896 NaN NaN NaN 1.791642 NaN NaN 2.199454
1 1981-01-05 1.548452 NaN NaN NaN 1.791642 NaN NaN 2.163397
2 1981-01-06 1.600452 NaN NaN NaN 1.791642 NaN NaN 2.199454
... ... ... ... ... ... ... ... ... ...
10327 2021-10-28 28.600000 7430.4 8.97 106.86 29.450000 12.10 12.02 26.230000
10328 2021-10-29 28.140000 7323.7 8.80 104.68 28.710000 11.83 11.72 25.670000
10329 2021-11-01 27.900000 7357.4 8.79 105.71 28.565000 12.03 11.83 24.050000

10330 rows × 9 columns

Plot

stocks.plot()

Data types and NA values

stocks.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 10330 entries, 0 to 10329
Data columns (total 9 columns):
 #   Column  Non-Null Count  Dtype  
---  ------  --------------  -----  
 0   Date    10330 non-null  object 
 1   ANZ     10319 non-null  float64
 2   ASX200  7452 non-null   float64
 3   BOQ     8970 non-null   float64
 4   CBA     7624 non-null   float64
 5   NAB     10316 non-null  float64
 6   QBE     9441 non-null   float64
 7   SUN     8424 non-null   float64
 8   WBC     10323 non-null  float64
dtypes: float64(8), object(1)
memory usage: 726.5+ KB
for col in stocks.columns:
    print(f"{col}: {stocks[col].isna().sum()}")
Date: 0
ANZ: 11
ASX200: 2878
BOQ: 1360
CBA: 2706
NAB: 14
QBE: 889
SUN: 1906
WBC: 7
asx200 = stocks.pop("ASX200")

Set the index to the date

stocks["Date"] = pd.to_datetime(stocks["Date"])
stocks = stocks.set_index("Date") # or `stocks.set_index("Date", inplace=True)`
stocks
ANZ BOQ CBA NAB QBE SUN WBC
Date
1981-01-02 1.588896 NaN NaN 1.791642 NaN NaN 2.199454
1981-01-05 1.548452 NaN NaN 1.791642 NaN NaN 2.163397
1981-01-06 1.600452 NaN NaN 1.791642 NaN NaN 2.199454
... ... ... ... ... ... ... ...
2021-10-28 28.600000 8.97 106.86 29.450000 12.10 12.02 26.230000
2021-10-29 28.140000 8.80 104.68 28.710000 11.83 11.72 25.670000
2021-11-01 27.900000 8.79 105.71 28.565000 12.03 11.83 24.050000

10330 rows × 7 columns

Plot II

stocks.plot()
plt.legend(loc="upper center", bbox_to_anchor=(0.5, -0.5), ncol=4);

Can index using dates I

stocks.loc["2010-1-4":"2010-01-8"]
ANZ BOQ CBA NAB QBE SUN WBC
Date
2010-01-04 22.89 10.772147 54.573702 26.046571 25.21 8.142453 25.012620
2010-01-05 23.00 10.910369 55.399220 26.379283 25.34 8.264684 25.220235
2010-01-06 22.66 10.855080 55.677708 25.865956 24.95 8.086039 25.101598
2010-01-07 22.12 10.523346 55.140624 25.656823 24.50 8.198867 24.765460
2010-01-08 22.25 10.781361 55.856736 25.571269 24.77 8.245879 24.864324

Note, these ranges are inclusive, not like Python’s normal slicing.

Can index using dates II

So to get 2019’s December and all of 2020 for CBA:

stocks.loc["2019-12":"2020", ["CBA"]]
CBA
Date
2019-12-02 81.43
2019-12-03 79.34
2019-12-04 77.81
... ...
2020-12-29 84.01
2020-12-30 83.59
2020-12-31 82.11

275 rows × 1 columns

Can look at the first differences

stocks.diff().plot()
plt.legend(loc="upper center", bbox_to_anchor=(0.5, -0.5), ncol=4);

Can look at the percentage changes

stocks.pct_change().plot()
plt.legend(loc="upper center", bbox_to_anchor=(0.5, -0.5), ncol=4);
/var/folders/y8/280cg1rd61jfcgvnkj769f7m0000gr/T/ipykernel_48082/1668655876.py:1: FutureWarning: The default fill_method='pad' in DataFrame.pct_change is deprecated and will be removed in a future version. Either fill in any non-leading NA values prior to calling pct_change or specify 'fill_method=None' to not fill NA values.
  stocks.pct_change().plot()

Focus on one stock

stock = stocks[["CBA"]]
stock
CBA
Date
1981-01-02 NaN
1981-01-05 NaN
1981-01-06 NaN
... ...
2021-10-28 106.86
2021-10-29 104.68
2021-11-01 105.71

10330 rows × 1 columns

stock.plot()

Find first non-missing value

first_day = stock.dropna().index[0]
first_day
Timestamp('1991-09-12 00:00:00')
stock = stock.loc[first_day:]
stock.isna().sum()
CBA    8
dtype: int64

Fill in the missing values

missing_day = stock[stock["CBA"].isna()].index[0]
prev_day = missing_day - pd.Timedelta(days=1)
after = missing_day + pd.Timedelta(days=3)
stock.loc[prev_day:after]
CBA
Date
2000-03-07 24.56662
2000-03-08 NaN
2000-03-09 NaN
2000-03-10 22.87580 
stock = stock.ffill()
stock.loc[prev_day:after]
CBA
Date
2000-03-07 24.56662
2000-03-08 24.56662
2000-03-09 24.56662
2000-03-10 22.87580 
stock.isna().sum()
CBA    0
dtype: int64

Baseline forecasts

Persistence forecast

The simplest model is to predict the next value to be the same as the current value.

stock.loc["2019":, "Persistence"] = stock.loc["2018"].iloc[-1].values[0]
stock.loc["2018-12":"2019"].plot()
plt.axvline("2019", color="black", linestyle="--")

Trend

We can extrapolate from recent trend:

past_date = stock.loc["2018"].index[-30]
past = stock.loc[past_date, "CBA"]
latest_date = stock.loc["2018", "CBA"].index[-1]
latest = stock.loc[latest_date, "CBA"]

trend = (latest - past) / (latest_date - past_date).days
print(trend)

tdays_since_cutoff = np.arange(1, len(stock.loc["2019":]) + 1)
stock.loc["2019":, "Trend"] = latest + trend * tdays_since_cutoff
0.07755555555555545

Trend forecasts

stock.loc["2018-12":"2019"].plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(ncol=3, loc="upper center", bbox_to_anchor=(0.5, 1.3))

Which is better?

If we look at the mean squared error (MSE) of the two models:

persistence_mse = mean_squared_error(stock.loc["2019", "CBA"], stock.loc["2019", "Persistence"])
trend_mse = mean_squared_error(stock.loc["2019", "CBA"], stock.loc["2019", "Trend"])
persistence_mse, trend_mse
(39.54629367588932, 37.87104674064297)

Use the history

cba_shifted = stock["CBA"].head().shift(1)
both = pd.concat([stock["CBA"].head(), cba_shifted], axis=1, keys=["Today", "Yesterday"])
both
Today Yesterday
Date
1991-09-12 6.425116 NaN
1991-09-13 6.365440 6.425116
1991-09-16 6.305764 6.365440
1991-09-17 6.285872 6.305764
1991-09-18 6.325656 6.285872 
def lagged_timeseries(df, target, window=30):
    lagged = pd.DataFrame()
    for i in range(window, 0, -1):
        lagged[f"T-{i}"] = df[target].shift(i)
    lagged["T"] = df[target].values
    return lagged

Lagged time series

df_lags = lagged_timeseries(stock, "CBA", 40)
df_lags
T-40 T-39 T-38 T-37 T-36 T-35 T-34 T-33 T-32 T-31 ... T-9 T-8 T-7 T-6 T-5 T-4 T-3 T-2 T-1 T
Date
1991-09-12 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 6.425116
1991-09-13 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN 6.425116 6.365440
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
2021-10-29 101.84 102.16 102.14 102.92 100.55 101.09 101.30 101.58 101.41 102.85 ... 103.94 103.89 105.03 104.95 104.88 105.46 105.1 106.10 106.860000 104.680000
2021-11-01 102.16 102.14 102.92 100.55 101.09 101.30 101.58 101.41 102.85 102.88 ... 103.89 105.03 104.95 104.88 105.46 105.10 106.1 106.86 104.680000 105.710000

7632 rows × 41 columns

Split into training and testing

# Split the data in time
X_train = df_lags.loc[:"2018"]
X_val = df_lags.loc["2019"]
X_test = df_lags.loc["2020":]

# Remove any with NAs and split into X and y
X_train = X_train.dropna()
X_val = X_val.dropna()
X_test = X_test.dropna()

y_train = X_train.pop("T")
y_val = X_val.pop("T")
y_test = X_test.pop("T")
X_train.shape, y_train.shape, X_val.shape, y_val.shape, X_test.shape, y_test.shape
((6872, 40), (6872,), (253, 40), (253,), (467, 40), (467,))

Inspect the split data

X_train
T-40 T-39 T-38 T-37 T-36 T-35 T-34 T-33 T-32 T-31 ... T-10 T-9 T-8 T-7 T-6 T-5 T-4 T-3 T-2 T-1
Date
1991-11-07 6.425116 6.365440 6.305764 6.285872 6.325656 6.385332 6.445008 6.445008 6.504684 6.564360 ... 7.280472 7.260580 7.190958 7.240688 7.379932 7.459500 7.320256 7.360040 7.459500 7.379932
1991-11-08 6.365440 6.305764 6.285872 6.325656 6.385332 6.445008 6.445008 6.504684 6.564360 6.624036 ... 7.260580 7.190958 7.240688 7.379932 7.459500 7.320256 7.360040 7.459500 7.379932 7.379932
1991-11-11 6.305764 6.285872 6.325656 6.385332 6.445008 6.445008 6.504684 6.564360 6.624036 6.663820 ... 7.190958 7.240688 7.379932 7.459500 7.320256 7.360040 7.459500 7.379932 7.379932 7.449554
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
2018-12-27 68.160000 69.230000 68.940000 68.350000 67.980000 68.950000 69.350000 70.620000 70.950000 71.770000 ... 68.430000 70.080000 70.180000 68.810000 69.260000 68.600000 69.400000 68.920000 68.480000 68.650000
2018-12-28 69.230000 68.940000 68.350000 67.980000 68.950000 69.350000 70.620000 70.950000 71.770000 70.900000 ... 70.080000 70.180000 68.810000 69.260000 68.600000 69.400000 68.920000 68.480000 68.650000 70.330000
2018-12-31 68.940000 68.350000 67.980000 68.950000 69.350000 70.620000 70.950000 71.770000 70.900000 69.210000 ... 70.180000 68.810000 69.260000 68.600000 69.400000 68.920000 68.480000 68.650000 70.330000 71.910000

6872 rows × 40 columns

Plot the split

Code 
y_train.plot()
y_val.plot()
y_test.plot()
plt.legend(["Train", "Validation", "Test"]);

Train on more recent data

X_train = X_train.loc["2012":]
y_train = y_train.loc["2012":]
Code 
y_train.plot()
y_val.plot()
y_test.plot()
plt.legend(["Train", "Validation", "Test"], loc="center left", bbox_to_anchor=(1, 0.5));

Rescale by eyeballing it

X_train = X_train / 100
X_val = X_val / 100
X_test = X_test / 100
y_train = y_train / 100
y_val = y_val / 100
y_test = y_test / 100
Code 
y_train.plot()
y_val.plot()
y_test.plot()
plt.legend(["Train", "Validation", "Test"], loc="center left", bbox_to_anchor=(1, 0.5));

Fit a linear model

lr = LinearRegression()
lr.fit(X_train, y_train);

Make a forecast for the validation data:

y_pred = lr.predict(X_val)
stock.loc[X_val.index, "Linear"] = y_pred
Code 
stock.loc["2018-12":"2019"].plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Inverse-transform the forecasts

stock.loc[X_val.index, "Linear"] = 100 * y_pred
Code 
stock.loc["2018-12":"2019"].plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Careful with the metrics

mean_squared_error(y_val, y_pred)
6.329105517812206e-05
mean_squared_error(100 * y_val, 100 * y_pred)
0.6329105517812207
100**2 * mean_squared_error(y_val, y_pred)
0.6329105517812206
linear_mse = 100**2 * mean_squared_error(y_val, y_pred)
persistence_mse, trend_mse, linear_mse
(39.54629367588932, 37.87104674064297, 0.6329105517812206)

Multi-step forecasts

Comparing apples to apples

The linear model is only producing one-step-ahead forecasts.

The other models are producing multi-step-ahead forecasts.

stock.loc["2019":, "Shifted"] = stock["CBA"].shift(1).loc["2019":]
Code 
stock.loc["2018-12":"2019"].plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

shifted_mse = mean_squared_error(stock.loc["2019", "CBA"], stock.loc["2019", "Shifted"])
persistence_mse, trend_mse, linear_mse, shifted_mse
(39.54629367588932, 37.87104674064297, 0.6329105517812206, 0.6367221343873524)

Autoregressive forecasts

The linear model needs the last 90 days to make a forecast.

Idea: Make the first forecast, then use that to make the next forecast, and so on.

\begin{aligned} \hat{y}_t &= \beta_0 + \beta_1 y_{t-1} + \beta_2 y_{t-2} + \ldots + \beta_n y_{t-n} \\ \hat{y}_{t+1} &= \beta_0 + \beta_1 \hat{y}_t + \beta_2 y_{t-1} + \ldots + \beta_n y_{t-n+1} \\ \hat{y}_{t+2} &= \beta_0 + \beta_1 \hat{y}_{t+1} + \beta_2 \hat{y}_t + \ldots + \beta_n y_{t-n+2} \end{aligned} \vdots \hat{y}_{t+k} = \beta_0 + \beta_1 \hat{y}_{t+k-1} + \beta_2 \hat{y}_{t+k-2} + \ldots + \beta_n \hat{y}_{t+k-n}

Autoregressive forecasting function

def autoregressive_forecast(model, X_val, suppress=False):
    """
    Generate a multi-step forecast using the given model.
    """
    multi_step = pd.Series(index=X_val.index, name="Multi Step")

    # Initialize the input data for forecasting
    input_data = X_val.iloc[0].values.reshape(1, -1)

    for i in range(len(multi_step)):
        # Ensure input_data has the correct feature names
        input_df = pd.DataFrame(input_data, columns=X_val.columns)
        if suppress:
            next_value = model.predict(input_df, verbose=0)
        else:
            next_value = model.predict(input_df) 

        multi_step.iloc[i] = next_value

        # Append that prediction to the input for the next forecast
        if i + 1 < len(multi_step):
            input_data = np.append(input_data[:, 1:], next_value).reshape(1, -1)

    return multi_step

Look at the autoregressive linear forecasts

lr_forecast = autoregressive_forecast(lr, X_val)
stock.loc[lr_forecast.index, "MS Linear"] = 100 * lr_forecast
stock.loc["2018-12":"2019"].drop(["Linear", "Shifted"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Metrics

One-step-ahead forecasts:

linear_mse, shifted_mse
(0.6329105517812206, 0.6367221343873524)

Multi-step-ahead forecasts:

multi_step_linear_mse = 100**2 * mean_squared_error(y_val, lr_forecast)
persistence_mse, trend_mse, multi_step_linear_mse
(39.54629367588932, 37.87104674064297, 23.84700379112976)

Prefer only short windows

stock.loc["2019":"2019-1"].drop(["Linear", "Shifted"], axis=1).plot();
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

“It’s tough to make predictions, especially about the future.”

Yogi Berra

Neural network forecasts

Simple feedforward neural network

model = Sequential([
        Dense(64, activation="leaky_relu"),
        Dense(1, "softplus")])

model.compile(optimizer="adam", loss="mean_squared_error")
if Path("aus_fin_fnn_model.h5").exists():
    model = keras.models.load_model("aus_fin_fnn_model.h5")
else:
    es = EarlyStopping(patience=15, restore_best_weights=True)
    model.fit(X_train, y_train, validation_data=(X_val, y_val), epochs=500,
        callbacks=[es], verbose=0)
    model.save("aus_fin_fnn_model.h5")

model.summary()
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Model: "sequential"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ dense (Dense)                   │ (32, 64)               │         2,624 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense_1 (Dense)                 │ (32, 1)                │            65 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 2,691 (10.52 KB)

 Trainable params: 2,689 (10.50 KB)

 Non-trainable params: 0 (0.00 B)

 Optimizer params: 2 (12.00 B)

Forecast and plot

y_pred = model.predict(X_val, verbose=0)
stock.loc[X_val.index, "FNN"] = 100 * y_pred
stock.loc["2018-12":"2019"].drop(["Persistence", "Trend", "MS Linear"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Autoregressive forecasts

nn_forecast = autoregressive_forecast(model, X_val, True)
stock.loc[nn_forecast.index, "MS FNN"] = 100 * nn_forecast
stock.loc["2018-12":"2019"].drop(["Linear", "Shifted", "FNN"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Metrics

One-step-ahead forecasts:

nn_mse = 100**2 * mean_squared_error(y_val, y_pred)
linear_mse, shifted_mse, nn_mse
(0.6329105517812206, 0.6367221343873524, 1.0445115967009995)

Multi-step-ahead forecasts:

multi_step_fnn_mse = 100**2 * mean_squared_error(y_val, nn_forecast)
persistence_mse, trend_mse, multi_step_linear_mse, multi_step_fnn_mse
(39.54629367588932, 37.87104674064297, 23.84700379112976, 10.150610173939775)

Recurrent Neural Networks

Basic facts of RNNs

  • A recurrent neural network is a type of neural network that is designed to process sequences of data (e.g. time series, sentences).
  • A recurrent neural network is any network that contains a recurrent layer.
  • A recurrent layer is a layer that processes data in a sequence.
  • An RNN can have one or more recurrent layers.
  • Weights are shared over time; this allows the model to be used on arbitrary-length sequences.

Applications

  • Forecasting: revenue forecast, weather forecast, predict disease rate from medical history, etc.
  • Classification: given a time series of the activities of a visitor on a website, classify whether the visitor is a bot or a human.
  • Event detection: given a continuous data stream, identify the occurrence of a specific event. Example: Detect utterances like “Hey Alexa” from an audio stream.
  • Anomaly detection: given a continuous data stream, detect anything unusual happening. Example: Detect unusual activity on the corporate network.

Origin of the name of RNNs

A recurrence relation is an equation that expresses each element of a sequence as a function of the preceding ones. More precisely, in the case where only the immediately preceding element is involved, a recurrence relation has the form

u_n = \psi(n, u_{n-1}) \quad \text{ for } \quad n > 0.

Example: Factorial n! = n (n-1)! for n > 0 given 0! = 1.

Source: Wikipedia, Recurrence relation.

Diagram of an RNN cell

The RNN processes each data in the sequence one by one, while keeping memory of what came before.

The following figure shows how the recurrent neural network combines an input X_l with a preprocessed state of the process A_l to produce the output O_l. RNNs have a cyclic information processing structure that enables them to pass information sequentially from previous inputs. RNNs can capture dependencies and patterns in sequential data, making them useful for analysing time series data.

Schematic of a recurrent neural network. E.g. SimpleRNN, LSTM, or GRU.

Source: James et al (2022), An Introduction to Statistical Learning, 2nd edition, Figure 10.12.

A SimpleRNN cell

Diagram of a SimpleRNN cell.

All the outputs before the final one are often discarded.

Source: Christopher Olah (2015), Understanding LSTM Networks, Colah’s Blog.

LSTM internals

Simple RNN structures encounter vanishing gradient problems, hence, struggle with learning long term dependencies. LSTM are designed to overcome this problem. LSTMs have a more complex network structure (contains more memory cells and gating mechanisms) and can better regulate the information flow.

Diagram of an LSTM cell. Notation for the diagram.

Source: Christopher Olah (2015), Understanding LSTM Networks, Colah’s Blog.

GRU internals

GRUs are simpler compared to LSTM, hence, computationally more efficient than LSTMs.

Diagram of a GRU cell.

Source: Christopher Olah (2015), Understanding LSTM Networks, Colah’s Blog.

Stock prediction with recurrent networks

SimpleRNN

from keras.layers import SimpleRNN, Reshape
model = Sequential([
        Reshape((-1, 1)),
        SimpleRNN(64, activation="tanh"),
        Dense(1, "softplus")])
model.compile(optimizer="adam", loss="mean_squared_error")
es = EarlyStopping(patience=15, restore_best_weights=True)
model.fit(X_train, y_train, validation_data=(X_val, y_val),
    epochs=500, callbacks=[es], verbose=0)
model.summary()
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Model: "sequential_1"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ reshape (Reshape)               │ (32, 40, 1)            │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ simple_rnn (SimpleRNN)          │ (32, 64)               │         4,224 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense_2 (Dense)                 │ (32, 1)                │            65 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 4,291 (16.77 KB)

 Trainable params: 4,289 (16.75 KB)

 Non-trainable params: 0 (0.00 B)

 Optimizer params: 2 (12.00 B)

Forecast and plot

y_pred = model.predict(X_val.to_numpy(), verbose=0)
stock.loc[X_val.index, "SimpleRNN"] = 100 * y_pred
Code 
stock.loc["2018-12":"2019"].drop(["Persistence", "Trend", "MS Linear", "MS FNN"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Multi-step forecasts

rnn_forecast = autoregressive_forecast(model, X_val, True)
stock.loc[rnn_forecast.index, "MS RNN"] = 100 * rnn_forecast
Code 
stock.loc["2018-12":"2019"].drop(["Linear", "Shifted", "FNN", "SimpleRNN"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Metrics

One-step-ahead forecasts:

rnn_mse = 100**2 * mean_squared_error(y_val, y_pred)
linear_mse, shifted_mse, nn_mse, rnn_mse
(0.6329105517812206,
 0.6367221343873524,
 1.0445115967009995,
 0.6444511280798111)

Multi-step-ahead forecasts:

multi_step_rnn_mse = 100**2 * mean_squared_error(y_val, rnn_forecast)
persistence_mse, trend_mse, multi_step_linear_mse, multi_step_fnn_mse, multi_step_rnn_mse
(39.54629367588932,
 37.87104674064297,
 23.84700379112976,
 10.150610173939775,
 10.583897540820221)

GRU

from keras.layers import GRU

model = Sequential([Reshape((-1, 1)),
        GRU(16, activation="tanh"),
        Dense(1, "softplus")])
model.compile(optimizer="adam", loss="mean_squared_error")
es = EarlyStopping(patience=15, restore_best_weights=True)
model.fit(X_train, y_train, validation_data=(X_val, y_val),
    epochs=500, callbacks=[es], verbose=0)
model.summary()
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Model: "sequential_2"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ reshape_1 (Reshape)             │ (32, 40, 1)            │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ gru (GRU)                       │ (32, 16)               │           912 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense_3 (Dense)                 │ (32, 1)                │            17 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 931 (3.64 KB)

 Trainable params: 929 (3.63 KB)

 Non-trainable params: 0 (0.00 B)

 Optimizer params: 2 (12.00 B)

Forecast and plot

y_pred = model.predict(X_val, verbose=0)
stock.loc[X_val.index, "GRU"] = 100 * y_pred
Code 
stock.loc["2018-12":"2019"].drop(["Persistence", "Trend", "MS Linear", "MS FNN", "MS RNN"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Multi-step forecasts

gru_forecast = autoregressive_forecast(model, X_val, True)
stock.loc[gru_forecast.index, "MS GRU"] = 100 * gru_forecast
Code 
stock.loc["2018-12":"2019"].drop(["Linear", "Shifted", "FNN", "SimpleRNN", "GRU"], axis=1).plot()
plt.axvline("2019", color="black", linestyle="--")
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5));

Metrics

One-step-ahead forecasts:

gru_mse = 100**2 * mean_squared_error(y_val, y_pred)
linear_mse, shifted_mse, nn_mse, rnn_mse, gru_mse
(0.6329105517812206,
 0.6367221343873524,
 1.0445115967009995,
 0.6444511280798111,
 0.63902721749169)

Multi-step-ahead forecasts:

multi_step_gru_mse = 100**2 * mean_squared_error(y_val, gru_forecast)
persistence_mse, trend_mse, multi_step_linear_mse, multi_step_fnn_mse, multi_step_rnn_mse, multi_step_gru_mse
(39.54629367588932,
 37.87104674064297,
 23.84700379112976,
 10.150610173939775,
 10.583897540820221,
 8.1113961776289)

Internals of the SimpleRNN

The rank of a time series

Say we had n observations of a time series x_1, x_2, \dots, x_n.

This \boldsymbol{x} = (x_1, \dots, x_n) would have shape (n,) & rank 1.

If instead we had a batch of b time series’

\boldsymbol{X} = \begin{pmatrix} x_7 & x_8 & \dots & x_{7+n-1} \\ x_2 & x_3 & \dots & x_{2+n-1} \\ \vdots & \vdots & \ddots & \vdots \\ x_3 & x_4 & \dots & x_{3+n-1} \\ \end{pmatrix} \,,

the batch \boldsymbol{X} would have shape (b, n) & rank 2.

Multivariate time series

Multivariate time series consists of more than 1 variable observation at a given time point. Following example has two variables x and y.

t x y
0 x_0 y_0
1 x_1 y_1
2 x_2 y_2
3 x_3 y_3

Say n observations of the m time series, would be a shape (n, m) matrix of rank 2.

In Keras, a batch of b of these time series has shape (b, n, m) and has rank 3.

Note

Use \boldsymbol{x}_t \in \mathbb{R}^{1 \times m} to denote the vector of all time series at time t. Here, \boldsymbol{x}_t = (x_t, y_t).

SimpleRNN

Say each prediction is a vector of size d, so \boldsymbol{y}_t \in \mathbb{R}^{1 \times d}.

Then the main equation of a SimpleRNN, given \boldsymbol{y}_0 = \boldsymbol{0}, is

\boldsymbol{y}_t = \psi\bigl( \boldsymbol{x}_t \boldsymbol{W}_x + \boldsymbol{y}_{t-1} \boldsymbol{W}_y + \boldsymbol{b} \bigr) .

Here, \begin{aligned} &\boldsymbol{x}_t \in \mathbb{R}^{1 \times m}, \boldsymbol{W}_x \in \mathbb{R}^{m \times d}, \\ &\boldsymbol{y}_{t-1} \in \mathbb{R}^{1 \times d}, \boldsymbol{W}_y \in \mathbb{R}^{d \times d}, \text{ and } \boldsymbol{b} \in \mathbb{R}^{d}. \end{aligned}

At each time step, a simple Recurrent Neural Network (RNN) takes an input vector x_t, incorporate it with the information from the previous hidden state {y}_{t-1} and produces an output vector at each time step y_t. The hidden state helps the network remember the context of the previous words, enabling it to make informed predictions about what comes next in the sequence. In a simple RNN, the output at time (t-1) is the same as the hidden state at time t.

SimpleRNN (in batches)

The difference between RNN and RNNs with batch processing lies in the way how the neural network handles sequences of input data. With batch processing, the model processes multiple (b) input sequences simultaneously. The training data is grouped into batches, and the weights are updated based on the average error across the entire batch. Batch processing often results in more stable weight updates, as the model learns from a diverse set of examples in each batch, reducing the impact of noise in individual sequences.

Say we operate on batches of size b, then \boldsymbol{Y}_t \in \mathbb{R}^{b \times d}.

The main equation of a SimpleRNN, given \boldsymbol{Y}_0 = \boldsymbol{0}, is \boldsymbol{Y}_t = \psi\bigl( \boldsymbol{X}_t \boldsymbol{W}_x + \boldsymbol{Y}_{t-1} \boldsymbol{W}_y + \boldsymbol{b} \bigr) . Here, \begin{aligned} &\boldsymbol{X}_t \in \mathbb{R}^{b \times m}, \boldsymbol{W}_x \in \mathbb{R}^{m \times d}, \\ &\boldsymbol{Y}_{t-1} \in \mathbb{R}^{b \times d}, \boldsymbol{W}_y \in \mathbb{R}^{d \times d}, \text{ and } \boldsymbol{b} \in \mathbb{R}^{d}. \end{aligned}

Remember, \boldsymbol{X} \in \mathbb{R}^{b \times n \times m}, \boldsymbol{Y} \in \mathbb{R}^{b \times d}, and \boldsymbol{X}_t is equivalent to X[:, t, :].

Simple Keras demo

1num_obs = 4
2num_time_steps = 3
3num_time_series = 2

X = (
    np.arange(num_obs * num_time_steps * num_time_series)
    .astype(np.float32)
    .reshape([num_obs, num_time_steps, num_time_series])
4)

output_size = 1
y = np.array([0, 0, 1, 1])
1
Defines the number of observations
2
Defines the number of time steps
3
Defines the number of time series
4
Reshapes the array to a range 3 tensor (4,3,2) 
1X[:2]
1
Selects the first two slices along the first dimension. Since the tensor of dimensions (4,3,2), X[:2] selects the first two slices (0 and 1) along the first dimension, and returns a sub-tensor of shape (2,3,2). 
array([[[ 0.,  1.],
        [ 2.,  3.],
        [ 4.,  5.]],

       [[ 6.,  7.],
        [ 8.,  9.],
        [10., 11.]]], dtype=float32)
1X[2:]
1
Selects the last two slices along the first dimension. The first dimension (axis=0) has size 4. Therefore, X[2:] selects the last two slices (2 and 3) along the first dimension, and returns a sub-tensor of shape (2,3,2). 
array([[[12., 13.],
        [14., 15.],
        [16., 17.]],

       [[18., 19.],
        [20., 21.],
        [22., 23.]]], dtype=float32)

Keras’ SimpleRNN

As usual, the SimpleRNN is just a layer in Keras.

1from keras.layers import SimpleRNN

2random.seed(1234)
3model = Sequential([SimpleRNN(output_size, activation="sigmoid")])
4model.compile(loss="binary_crossentropy", metrics=["accuracy"])

5hist = model.fit(X, y, epochs=500, verbose=False)
6model.evaluate(X, y, verbose=False)
1
Imports the SimpleRNN layer from the Keras library
2
Sets the seed for the random number generator to ensure reproducibility
3
Defines a simple RNN with one output node and sigmoid activation function
4
Specifies binary crossentropy as the loss function (usually used in classification problems), and specifies “accuracy” as the metric to be monitored during training
5
Trains the model for 500 epochs and saves output as hist
6
Evaluates the model to obtain a value for the loss and accuracy 
[3.1487133502960205, 0.5]

The predicted probabilities on the training set are:

model.predict(X, verbose=0)
array([[0.97],
       [1.  ],
       [1.  ],
       [1.  ]], dtype=float32)

SimpleRNN weights

To verify the results of predicted probabilities, we can obtain the weights of the fitted model and calculate the outcome manually as follows.

model.get_weights()
[array([[0.68],
        [0.2 ]], dtype=float32),
 array([[0.49]], dtype=float32),
 array([-0.51], dtype=float32)]
def sigmoid(x):
    return 1 / (1 + np.exp(-x))


W_x, W_y, b = model.get_weights()

Y = np.zeros((num_obs, output_size), dtype=np.float32)
for t in range(num_time_steps):
    X_t = X[:, t, :]
    z = X_t @ W_x + Y @ W_y + b
    Y = sigmoid(z)

Y
array([[0.97],
       [1.  ],
       [1.  ],
       [1.  ]], dtype=float32)

Other recurrent network variants

Input and output sequences

Categories of recurrent neural networks: sequence to sequence, sequence to vector, vector to sequence, encoder-decoder network.

Source: Aurélien Géron (2019), Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 2nd Edition, Chapter 15.

Input and output sequences

  • Sequence to sequence: Useful for predicting time series such as using prices over the last N days to output the prices shifted one day into the future (i.e. from N-1 days ago to tomorrow.)
  • Sequence to vector: ignore all outputs in the previous time steps except for the last one. Example: give a sentiment score to a sequence of words corresponding to a movie review.

Input and output sequences

  • Vector to sequence: feed the network the same input vector over and over at each time step and let it output a sequence. Example: given that the input is an image, find a caption for it. The image is treated as an input vector (pixels in an image do not follow a sequence). The caption is a sequence of textual description of the image. A dataset containing images and their descriptions is the input of the RNN.
  • The Encoder-Decoder: The encoder is a sequence-to-vector network. The decoder is a vector-to-sequence network. Example: Feed the network a sequence in one language. Use the encoder to convert the sentence into a single vector representation. The decoder decodes this vector into the translation of the sentence in another language.

Recurrent layers can be stacked.

Deep RNN unrolled through time.

Source: Aurélien Géron (2019), Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 2nd Edition, Chapter 15.

CoreLogic Hedonic Home Value Index

Australian House Price Indices

Note

I apologise in advance for not being able to share this dataset with anyone (it is not mine to share).

Percentage changes

changes = house_prices.pct_change().dropna()
changes.round(2)
Brisbane East_Bris North_Bris West_Bris Melbourne North_Syd Sydney
Date
1990-02-28 0.03 -0.01 0.01 0.01 0.00 -0.00 -0.02
1990-03-31 0.01 0.03 0.01 0.01 0.02 -0.00 0.03
1990-04-30 0.02 0.02 0.01 -0.00 0.01 0.03 0.04
... ... ... ... ... ... ... ...
2021-03-31 0.04 0.04 0.03 0.04 0.02 0.05 0.05
2021-04-30 0.03 0.01 0.01 -0.00 0.01 0.02 0.02
2021-05-31 0.03 0.03 0.03 0.03 0.03 0.02 0.04

376 rows × 7 columns

Percentage changes

changes.plot();

The size of the changes

changes.mean()
Brisbane      0.005496
East_Bris     0.005416
North_Bris    0.005024
West_Bris     0.004842
Melbourne     0.005677
North_Syd     0.004819
Sydney        0.005526
dtype: float64
changes *= 100
changes.mean()
Brisbane      0.549605
East_Bris     0.541562
North_Bris    0.502390
West_Bris     0.484204
Melbourne     0.567700
North_Syd     0.481863
Sydney        0.552641
dtype: float64
changes.plot(legend=False);

Split without shuffling

num_train = int(0.6 * len(changes))
num_val = int(0.2 * len(changes))
num_test = len(changes) - num_train - num_val
print(f"# Train: {num_train}, # Val: {num_val}, # Test: {num_test}")
# Train: 225, # Val: 75, # Test: 76

Subsequences of a time series

Keras has a built-in method for converting a time series into subsequences/chunks.

from keras.utils import timeseries_dataset_from_array

integers = range(10)
dummy_dataset = timeseries_dataset_from_array(
    data=integers[:-3],
    targets=integers[3:],
    sequence_length=3,
    batch_size=2,
)

for inputs, targets in dummy_dataset:
    for i in range(inputs.shape[0]):
        print([int(x) for x in inputs[i]], int(targets[i]))
[0, 1, 2] 3
[1, 2, 3] 4
[2, 3, 4] 5
[3, 4, 5] 6
[4, 5, 6] 7

Source: Code snippet in Chapter 10 of Chollet.

Predicting Sydney House Prices

Creating dataset objects

# Num. of input time series.
num_ts = changes.shape[1]

# How many prev. months to use.
seq_length = 6

# Predict the next month ahead.
ahead = 1

# The index of the first target.
delay = seq_length + ahead - 1
# Which suburb to predict.
target_suburb = changes["Sydney"]

train_ds = timeseries_dataset_from_array(
    changes[:-delay],
    targets=target_suburb[delay:],
    sequence_length=seq_length,
    end_index=num_train,
)
val_ds = timeseries_dataset_from_array(
    changes[:-delay],
    targets=target_suburb[delay:],
    sequence_length=seq_length,
    start_index=num_train,
    end_index=num_train + num_val,
)
test_ds = timeseries_dataset_from_array(
    changes[:-delay],
    targets=target_suburb[delay:],
    sequence_length=seq_length,
    start_index=num_train + num_val,
)

Converting Dataset to numpy

The Dataset object can be handed to Keras directly, but if we really need a numpy array, we can run:

X_train = np.concatenate(list(train_ds.map(lambda x, y: x)))
y_train = np.concatenate(list(train_ds.map(lambda x, y: y)))
2025-04-01 10:04:47.797327: I tensorflow/core/framework/local_rendezvous.cc:405] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence

The shape of our training set is now:

X_train.shape
(220, 6, 7)
y_train.shape
(220,)

Converting the rest to numpy arrays:

X_val = np.concatenate(list(val_ds.map(lambda x, y: x)))
y_val = np.concatenate(list(val_ds.map(lambda x, y: y)))
X_test = np.concatenate(list(test_ds.map(lambda x, y: x)))
y_test = np.concatenate(list(test_ds.map(lambda x, y: y)))
2025-04-01 10:04:48.173016: I tensorflow/core/framework/local_rendezvous.cc:405] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence

A dense network

from keras.layers import Input, Flatten
random.seed(1)
model_dense = Sequential([
    Input((seq_length, num_ts)),
    Flatten(),
    Dense(50, activation="leaky_relu"),
    Dense(20, activation="leaky_relu"),
    Dense(1, activation="linear")
])
model_dense.compile(loss="mse", optimizer="adam")
print(f"This model has {model_dense.count_params()} parameters.")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)
%time hist = model_dense.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0);
This model has 3191 parameters.
Epoch 57: early stopping
Restoring model weights from the end of the best epoch: 7.
CPU times: user 3.19 s, sys: 410 ms, total: 3.6 s
Wall time: 15.9 s

Plot the model

from keras.utils import plot_model

plot_model(model_dense, show_shapes=True)

Assess the fits

model_dense.evaluate(X_val, y_val, verbose=0)
1.1644607782363892
Code 
y_pred = model_dense.predict(X_val, verbose=0)
plt.plot(y_val, label="Sydney")
plt.plot(y_pred, label="Dense")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

A SimpleRNN layer

random.seed(1)

model_simple = Sequential([
    Input((seq_length, num_ts)),
    SimpleRNN(50),
    Dense(1, activation="linear")
])
model_simple.compile(loss="mse", optimizer="adam")
print(f"This model has {model_simple.count_params()} parameters.")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)
%time hist = model_simple.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0);
This model has 2951 parameters.
Epoch 62: early stopping
Restoring model weights from the end of the best epoch: 12.
CPU times: user 3.77 s, sys: 388 ms, total: 4.16 s
Wall time: 15.4 s

Plot the model

plot_model(model_simple, show_shapes=True)

Assess the fits

model_simple.evaluate(X_val, y_val, verbose=0)
1.2507916688919067
Code 
y_pred = model_simple.predict(X_val, verbose=0)

plt.plot(y_val, label="Sydney")
plt.plot(y_pred, label="SimpleRNN")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);
WARNING:tensorflow:5 out of the last 259 calls to <function TensorFlowTrainer.make_predict_function.<locals>.one_step_on_data_distributed at 0x368a77920> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has reduce_retracing=True option that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
WARNING:tensorflow:5 out of the last 259 calls to <function TensorFlowTrainer.make_predict_function.<locals>.one_step_on_data_distributed at 0x368a77920> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has reduce_retracing=True option that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
WARNING:tensorflow:6 out of the last 261 calls to <function TensorFlowTrainer.make_predict_function.<locals>.one_step_on_data_distributed at 0x368a77920> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has reduce_retracing=True option that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
WARNING:tensorflow:6 out of the last 261 calls to <function TensorFlowTrainer.make_predict_function.<locals>.one_step_on_data_distributed at 0x368a77920> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has reduce_retracing=True option that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.

A LSTM layer

from keras.layers import LSTM

random.seed(1)

model_lstm = Sequential([
    Input((seq_length, num_ts)),
    LSTM(50),
    Dense(1, activation="linear")
])

model_lstm.compile(loss="mse", optimizer="adam")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)

%time hist = model_lstm.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0);
Epoch 59: early stopping
Restoring model weights from the end of the best epoch: 9.
CPU times: user 4.21 s, sys: 415 ms, total: 4.62 s
Wall time: 14.9 s

Assess the fits

model_lstm.evaluate(X_val, y_val, verbose=0)
0.8353261947631836
Code 
y_pred = model_lstm.predict(X_val, verbose=0)
plt.plot(y_val, label="Sydney")
plt.plot(y_pred, label="LSTM")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

A GRU layer

from keras.layers import GRU

random.seed(1)

model_gru = Sequential([
    Input((seq_length, num_ts)),
    GRU(50),
    Dense(1, activation="linear")
])

model_gru.compile(loss="mse", optimizer="adam")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)

%time hist = model_gru.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0)
Epoch 57: early stopping
Restoring model weights from the end of the best epoch: 7.
CPU times: user 4.38 s, sys: 447 ms, total: 4.82 s
Wall time: 14.5 s

Assess the fits

model_gru.evaluate(X_val, y_val, verbose=0)
0.7435100078582764
Code 
y_pred = model_gru.predict(X_val, verbose=0)
plt.plot(y_val, label="Sydney")
plt.plot(y_pred, label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Two GRU layers

random.seed(1)

model_two_grus = Sequential([
    Input((seq_length, num_ts)),
    GRU(50, return_sequences=True),
    GRU(50),
    Dense(1, activation="linear")
])

model_two_grus.compile(loss="mse", optimizer="adam")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)

%time hist = model_two_grus.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0)
Epoch 56: early stopping
Restoring model weights from the end of the best epoch: 6.
CPU times: user 5.45 s, sys: 507 ms, total: 5.96 s
Wall time: 16 s

Assess the fits

model_two_grus.evaluate(X_val, y_val, verbose=0)
0.7989509105682373
Code 
y_pred = model_two_grus.predict(X_val, verbose=0)
plt.plot(y_val, label="Sydney")
plt.plot(y_pred, label="2 GRUs")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Compare the models

Model MSE
1 SimpleRNN 1.250792
0 Dense 1.164461
2 LSTM 0.835326
4 2 GRUs 0.798951
3 GRU 0.743510

The network with two GRU layers is the best.

model_two_grus.evaluate(X_test, y_test, verbose=0)
1.855254888534546

Test set

Code 
y_pred = model_two_grus.predict(X_test, verbose=0)
plt.plot(y_test, label="Sydney")
plt.plot(y_pred, label="2 GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Predicting Multiple Time Series

Creating dataset objects

Change the targets argument to include all the suburbs.

train_ds = timeseries_dataset_from_array(
    changes[:-delay],
    targets=changes[delay:],
    sequence_length=seq_length,
    end_index=num_train,
)
val_ds = timeseries_dataset_from_array(
    changes[:-delay],
    targets=changes[delay:],
    sequence_length=seq_length,
    start_index=num_train,
    end_index=num_train + num_val,
)
test_ds = timeseries_dataset_from_array(
    changes[:-delay],
    targets=changes[delay:],
    sequence_length=seq_length,
    start_index=num_train + num_val,
)

Converting Dataset to numpy

The shape of our training set is now:

X_train = np.concatenate(list(train_ds.map(lambda x, y: x)))
X_train.shape
2025-04-01 10:06:15.319772: I tensorflow/core/framework/local_rendezvous.cc:405] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence
(220, 6, 7)
y_train = np.concatenate(list(train_ds.map(lambda x, y: y)))
y_train.shape
(220, 7)

Converting the rest to numpy arrays:

X_val = np.concatenate(list(val_ds.map(lambda x, y: x)))
y_val = np.concatenate(list(val_ds.map(lambda x, y: y)))
X_test = np.concatenate(list(test_ds.map(lambda x, y: x)))
y_test = np.concatenate(list(test_ds.map(lambda x, y: y)))

A dense network

random.seed(1)
model_dense = Sequential([
    Input((seq_length, num_ts)),
    Flatten(),
    Dense(50, activation="leaky_relu"),
    Dense(20, activation="leaky_relu"),
    Dense(num_ts, activation="linear")
])
model_dense.compile(loss="mse", optimizer="adam")
print(f"This model has {model_dense.count_params()} parameters.")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)
%time hist = model_dense.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0);
This model has 3317 parameters.
Epoch 75: early stopping
Restoring model weights from the end of the best epoch: 25.
CPU times: user 3.89 s, sys: 482 ms, total: 4.37 s
Wall time: 16 s

Plot the model

plot_model(model_dense, show_shapes=True)

Assess the fits

model_dense.evaluate(X_val, y_val, verbose=0)
1.4294650554656982
Code 
Y_pred = model_dense.predict(X_val, verbose=0)
plt.scatter(y_val, Y_pred)
add_diagonal_line()
plt.xlabel("Actual")
plt.ylabel("Predicted")
plt.show()

plt.plot(y_val[:, 4], label="Melbourne")
plt.plot(Y_pred[:, 4], label="Dense")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Code 
plt.plot(y_val[:, 0], label="Brisbane")
plt.plot(Y_pred[:, 0], label="Dense")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False)
plt.show()

plt.plot(y_val[:, 6], label="Sydney")
plt.plot(Y_pred[:, 6], label="Dense")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

A SimpleRNN layer

random.seed(1)

model_simple = Sequential([
    Input((seq_length, num_ts)),
    SimpleRNN(50),
    Dense(num_ts, activation="linear")
])
model_simple.compile(loss="mse", optimizer="adam")
print(f"This model has {model_simple.count_params()} parameters.")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)
%time hist = model_simple.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0);
This model has 3257 parameters.
Epoch 70: early stopping
Restoring model weights from the end of the best epoch: 20.
CPU times: user 4.15 s, sys: 480 ms, total: 4.63 s
Wall time: 15.6 s

Plot the model

plot_model(model_simple, show_shapes=True)

Assess the fits

model_simple.evaluate(X_val, y_val, verbose=0)
1.4916820526123047
Code 
Y_pred = model_simple.predict(X_val, verbose=0)
plt.scatter(y_val, Y_pred)
add_diagonal_line()
plt.xlabel("Actual")
plt.ylabel("Predicted")
plt.show()

plt.plot(y_val[:, 4], label="Melbourne")
plt.plot(Y_pred[:, 4], label="SimpleRNN")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Code 
plt.plot(y_val[:, 0], label="Brisbane")
plt.plot(Y_pred[:, 0], label="SimpleRNN")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False)
plt.show()

plt.plot(y_val[:, 6], label="Sydney")
plt.plot(Y_pred[:, 6], label="SimpleRNN")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

A LSTM layer

random.seed(1)

model_lstm = Sequential([
    Input((seq_length, num_ts)),
    LSTM(50),
    Dense(num_ts, activation="linear")
])

model_lstm.compile(loss="mse", optimizer="adam")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)

%time hist = model_lstm.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0);
Epoch 74: early stopping
Restoring model weights from the end of the best epoch: 24.
CPU times: user 4.8 s, sys: 572 ms, total: 5.37 s
Wall time: 16.2 s

Assess the fits

model_lstm.evaluate(X_val, y_val, verbose=0)
1.3311247825622559
Code 
Y_pred = model_lstm.predict(X_val, verbose=0)
plt.scatter(y_val, Y_pred)
add_diagonal_line()
plt.xlabel("Actual")
plt.ylabel("Predicted")
plt.show()

plt.plot(y_val[:, 4], label="Melbourne")
plt.plot(Y_pred[:, 4], label="LSTM")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Code 
plt.plot(y_val[:, 0], label="Brisbane")
plt.plot(Y_pred[:, 0], label="LSTM")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False)
plt.show()

plt.plot(y_val[:, 6], label="Sydney")
plt.plot(Y_pred[:, 6], label="LSTM")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

A GRU layer

random.seed(1)

model_gru = Sequential([
    Input((seq_length, num_ts)),
    GRU(50),
    Dense(num_ts, activation="linear")
])

model_gru.compile(loss="mse", optimizer="adam")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)

%time hist = model_gru.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0)
Epoch 70: early stopping
Restoring model weights from the end of the best epoch: 20.
CPU times: user 5.04 s, sys: 543 ms, total: 5.58 s
Wall time: 15.7 s

Assess the fits

model_gru.evaluate(X_val, y_val, verbose=0)
1.344503402709961
Code 
Y_pred = model_gru.predict(X_val, verbose=0)
plt.scatter(y_val, Y_pred)
add_diagonal_line()
plt.xlabel("Actual")
plt.ylabel("Predicted")
plt.show()

plt.plot(y_val[:, 4], label="Melbourne")
plt.plot(Y_pred[:, 4], label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Code 
plt.plot(y_val[:, 0], label="Brisbane")
plt.plot(Y_pred[:, 0], label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False)
plt.show()

plt.plot(y_val[:, 6], label="Sydney")
plt.plot(Y_pred[:, 6], label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Two GRU layers

random.seed(1)

model_two_grus = Sequential([
    Input((seq_length, num_ts)),
    GRU(50, return_sequences=True),
    GRU(50),
    Dense(num_ts, activation="linear")
])

model_two_grus.compile(loss="mse", optimizer="adam")

es = EarlyStopping(patience=50, restore_best_weights=True, verbose=1)

%time hist = model_two_grus.fit(X_train, y_train, epochs=1_000, \
  validation_data=(X_val, y_val), callbacks=[es], verbose=0)
Epoch 67: early stopping
Restoring model weights from the end of the best epoch: 17.
CPU times: user 6.19 s, sys: 612 ms, total: 6.8 s
Wall time: 18.1 s

Assess the fits

model_two_grus.evaluate(X_val, y_val, verbose=0)
1.358651041984558
Code 
Y_pred = model_two_grus.predict(X_val, verbose=0)
plt.scatter(y_val, Y_pred)
add_diagonal_line()
plt.xlabel("Actual")
plt.ylabel("Predicted")
plt.show()

plt.plot(y_val[:, 4], label="Melbourne")
plt.plot(Y_pred[:, 4], label="2 GRUs")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Code 
plt.plot(y_val[:, 0], label="Brisbane")
plt.plot(Y_pred[:, 0], label="2 GRUs")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False)
plt.show()

plt.plot(y_val[:, 6], label="Sydney")
plt.plot(Y_pred[:, 6], label="2 GRUs")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Compare the models

Code 
models = [model_dense, model_simple, model_lstm, model_gru, model_two_grus]
model_names = ["Dense", "SimpleRNN", "LSTM", "GRU", "2 GRUs"]
mse_val = {
    name: model.evaluate(X_val, y_val, verbose=0)
    for name, model in zip(model_names, models)
}
val_results = pd.DataFrame({"Model": mse_val.keys(), "MSE": mse_val.values()})
val_results.sort_values("MSE", ascending=False)
Model MSE
1 SimpleRNN 1.491682
0 Dense 1.429465
4 2 GRUs 1.358651
3 GRU 1.344503
2 LSTM 1.331125

The network with an LSTM layer is the best.

model_lstm.evaluate(X_test, y_test, verbose=0)
1.932026982307434

Test set

Code 
Y_pred = model_lstm.predict(X_test, verbose=0)
plt.scatter(y_test, Y_pred)
add_diagonal_line()
plt.xlabel("Actual")
plt.ylabel("Predicted")
plt.show()

plt.plot(y_test[:, 4], label="Melbourne")
plt.plot(Y_pred[:, 4], label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Code 
plt.plot(y_test[:, 0], label="Brisbane")
plt.plot(Y_pred[:, 0], label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False)
plt.show()

plt.plot(y_test[:, 6], label="Sydney")
plt.plot(Y_pred[:, 6], label="GRU")
plt.xlabel("Time")
plt.ylabel("Change in HPI (%)")
plt.legend(frameon=False);

Package Versions

from watermark import watermark
print(watermark(python=True, packages="keras,matplotlib,numpy,pandas,seaborn,scipy,torch,tensorflow,tf_keras"))
Python implementation: CPython
Python version       : 3.11.11
IPython version      : 8.32.0

keras     : 3.8.0
matplotlib: 3.10.0
numpy     : 1.26.4
pandas    : 2.2.3
seaborn   : 0.13.2
scipy     : 1.13.1
torch     : 2.5.1
tensorflow: 2.18.0
tf_keras  : 2.18.0

Glossary

  • autoregressive forecasting
  • forecasting
  • GRU
  • LSTM
  • one-step/multi-step ahead forecasting
  • persistence forecast
  • recurrent neural networks
  • SimpleRNN